Large earthquakes often disrupt road networks, severely hindering the timely delivery of emergency supplies. This paper studies a truck–UAV collaborative emergency routing problem under road blockage conditions. We develop a mixed-integer programming model that coordinates truck and UAV operations to minimize total emergency response time while accounting for payload, endurance, demand satisfaction, and road repair constraints. To solve the problem efficiently, we propose an improved Variable Neighborhood Search algorithm with greedy initialization (VNS-G), together with a benchmark variant based on random initialization (VNS-R). Computational experiments on instances of different sizes are conducted to evaluate the proposed approach. The results show that VNS-G can obtain high-quality solutions close to those of CPLEX on small-scale instances. On medium-scale instances, it provides substantial computational savings while maintaining competitive solution quality, reducing computation time from 682–3496 s for CPLEX to 95–340 s in the tested cases. For large-scale instances, where exact optimization becomes computationally impractical, the proposed heuristic remains effective in generating feasible solutions within operationally meaningful time. A case study of the earthquake-prone Ya'an region further illustrates the model's practical applicability. Sensitivity analysis reveals a mechanism-based managerial insight: UAV endurance primarily affects reachability, whereas UAV payload more directly improves response efficiency by reducing the need for repeated sorties. This suggests that, once reachability is ensured, improving payload capacity is likely to generate greater operational benefits than further extending endurance.